# yolov10_realtime_endtoend_object_detection__caf129e5.pdf

YOLOv10: Real-Time End-to-End Object Detection

Ao Wang1 Hui Chen2 Lihao Liu1 Kai Chen1

Zijia Lin1 Jungong Han3 Guiguang Ding1

1School of Software, Tsinghua University 2BNRist, Tsinghua University 3Department of Automation, Tsinghua University wa22@mails.tsinghua.edu.cn huichen@mail.tsinghua.edu.cn linzijia07@tsinghua.org.cn {louisliu2048,chenkai2010.9,jungonghan77}@gmail.com dinggg@tsinghua.edu.cn

2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Latency (ms)

COCO AP (%)

YOLOv6-v3.0 YOLOv7 YOLOv8 YOLOv9 PPYOLOE RTMDet YOLO-MS Gold-YOLO RT-DETR YOLOv10 (Ours)

0 20 40 60 80 100 Number of Parameters (M)

COCO AP (%)

YOLOv6-v3.0 YOLOv7 YOLOv8 YOLOv9 PPYOLOE RTMDet YOLO-MS Gold-YOLO RT-DETR YOLOv10 (Ours)

Figure 1: Comparisons with others in terms of latency-accuracy (left) and size-accuracy (right) trade-offs. We measure the end-to-end latency using the official pre-trained models.

Over the past years, YOLOs have emerged as the predominant paradigm in the field of real-time object detection owing to their effective balance between computational cost and detection performance. Researchers have explored the architectural designs, optimization objectives, data augmentation strategies, and others for YOLOs, achieving notable progress. However, the reliance on the non-maximum suppression (NMS) for post-processing hampers the end-to-end deployment of YOLOs and adversely impacts the inference latency. Besides, the design of various components in YOLOs lacks the comprehensive and thorough inspection, resulting in noticeable computational redundancy and limiting the model s capability. It renders the suboptimal efficiency, along with considerable potential for performance improvements. In this work, we aim to further advance the performance-efficiency boundary of YOLOs from both the post-processing and the model architecture. To this end, we first present the consistent dual assignments for NMS-free training of YOLOs, which brings the competitive performance and low inference latency simultaneously. Moreover, we introduce the holistic efficiency-accuracy driven model design strategy for YOLOs. We comprehensively optimize various components of YOLOs from both the efficiency and accuracy perspectives, which greatly reduces the computational overhead and enhances the capability. The outcome of our effort is a new generation of YOLO series for real-time end-to-end object detection, dubbed YOLOv10. Extensive experiments show that YOLOv10 achieves the state-of-the-art performance and efficiency across various model scales. For

Corresponding author.

38th Conference on Neural Information Processing Systems (Neur IPS 2024).

example, our YOLOv10-S is 1.8 faster than RT-DETR-R18 under the similar AP on COCO, meanwhile enjoying 2.8 smaller number of parameters and FLOPs. Compared with YOLOv9-C, YOLOv10-B has 46% less latency and 25% fewer parameters for the same performance. Code and models are available at https://github.com/THU-MIG/yolov10.

1 Introduction

Real-time object detection has always been a focal point of research in the area of computer vision, which aims to accurately predict the categories and positions of objects in an image under low latency. It is widely adopted in various practical applications, including autonomous driving [3], robot navigation [12], and object tracking [72], etc. In recent years, researchers have concentrated on devising CNN-based object detectors to achieve real-time detection [19, 23, 48, 49, 50, 57, 13]. Among them, YOLOs have gained increasing popularity due to their adept balance between performance and efficiency [2, 20, 29, 20, 21, 65, 60, 70, 8, 71, 17, 29]. The detection pipeline of YOLOs consists of two parts: the model forward process and the NMS post-processing. However, both of them still have deficiencies, resulting in suboptimal accuracy-latency boundaries.

Specifically, YOLOs usually employ one-to-many label assignment strategy during training, whereby one ground-truth object corresponds to multiple positive samples. Despite yielding superior performance, this approach necessitates NMS to select the best positive prediction during inference. This slows down the inference speed and renders the performance sensitive to the hyperparameters of NMS, thereby preventing YOLOs from achieving optimal end-to-end deployment [78]. One line to tackle this issue is to adopt the recently introduced end-to-end DETR architectures [4, 81, 73, 30, 36, 42, 67]. For example, RT-DETR [78] presents an efficient hybrid encoder and uncertainty-minimal query selection, propelling DETRs into the realm of real-time applications. Nevertheless, when considering only the forward process of model during deployment, the efficiency of the DETRs still has room for improvements compared with YOLOs. Another line is to explore end-to-end detection for CNNbased detectors, which typically leverages one-to-one assignment strategies to suppress the redundant predictions [6, 55, 66, 80, 17]. However, they usually introduce additional inference overhead or achieve suboptimal performance for YOLOs.

Furthermore, the model architecture design remains a fundamental challenge for YOLOs, which exhibits an important impact on the accuracy and speed [50, 17, 71, 8]. To achieve more efficient and effective model architectures, researchers have explored different design strategies. Various primary computational units are presented for the backbone to enhance the feature extraction ability, including Dark Net [48, 49, 50], CSPNet [2], Efficient Rep [29] and ELAN [62, 64], etc. For the neck, PAN [37], Bi C [29], GD [60] and Rep GFPN [71], etc., are explored to enhance the multi-scale feature fusion. Besides, model scaling strategies [62, 61] and re-parameterization [11, 29] techniques are also investigated. While these efforts have achieved notable advancements, a comprehensive inspection for various components in YOLOs from both the efficiency and accuracy perspectives is still lacking. As a result, there still exists considerable computational redundancy within YOLOs, leading to inefficient parameter utilization and suboptimal efficiency. Besides, the resulting constrained model capability also leads to inferior performance, leaving ample room for accuracy improvements.

In this work, we aim to address these issues and further advance the accuracy-speed boundaries of YOLOs. We target both the post-processing and the model architecture throughout the detection pipeline. To this end, we first tackle the problem of redundant predictions in the post-processing by presenting a consistent dual assignments strategy for NMS-free YOLOs with the dual label assignments and consistent matching metric. It allows the model to enjoy rich and harmonious supervision during training while eliminating the need for NMS during inference, leading to competitive performance with high efficiency. Secondly, we propose the holistic efficiency-accuracy driven model design strategy for the model architecture by performing the comprehensive inspection for various components in YOLOs. For efficiency, we propose the lightweight classification head, spatial-channel decoupled downsampling, and rank-guided block design, to reduce the manifested computational redundancy and achieve more efficient architecture. For accuracy, we explore the large-kernel convolution and present the effective partial self-attention module to enhance the model capability, harnessing the potential for performance improvements under low cost.

Based on these approaches, we succeed in achieving a new family of real-time end-to-end detectors with different model scales, i.e., YOLOv10-N / S / M / B / L / X. Extensive experiments on standard benchmarks for object detection, i.e., COCO [35], demonstrate that our YOLOv10 can significantly outperform previous state-of-the-art models in terms of computation-accuracy trade-offs across various model scales. As shown in Fig. 1, our YOLOv10-S / X are 1.8 / 1.3 faster than RT-DETRR18 / R101, respectively, under the similar performance. Compared with YOLOv9-C, YOLOv10-B achieves a 46% reduction in latency with the same performance. Moreover, YOLOv10 exhibits highly efficient parameter utilization. Our YOLOv10-L / X outperforms YOLOv8-L / X by 0.3 AP and 0.5 AP, with 1.8 and 2.3 smaller number of parameters, respectively. YOLOv10-M achieves the similar AP compared with YOLOv9-M / YOLO-MS, with 23% / 31% fewer parameters, respectively. We hope that our work can inspire further studies and advancements in the field.

2 Related Work

Real-time object detectors. Real-time object detection aims to classify and locate objects under low latency, which is crucial for real-world applications. Over the past years, substantial efforts have been directed towards developing efficient detectors [19, 57, 48, 34, 79, 75, 32, 31, 41]. Particularly, the YOLO series [48, 49, 50, 2, 20, 29, 62, 21, 65] stand out as the mainstream ones. YOLOv1, YOLOv2, and YOLOv3 identify the typical detection architecture consisting of three parts, i.e., backbone, neck, and head [48, 49, 50]. YOLOv4 [2] and YOLOv5 [20] introduce the CSPNet [63] design to replace Dark Net [47], coupled with data augmentation strategies, enhanced PAN, and a greater variety of model scales, etc. YOLOv6 [29] presents Bi C and Sim CSPSPPF for neck and backbone, respectively, with anchor-aided training and self-distillation strategy. YOLOv7 [62] introduces E-ELAN for rich gradient flow path and explores several trainable bag-of-freebies methods. YOLOv8 [21] presents C2f building block for effective feature extraction and fusion. Gold-YOLO [60] provides the advanced GD mechanism to boost the multi-scale feature fusion capability. YOLOv9 [65] proposes GELAN to improve the architecture and PGI to augment the training process.

End-to-end object detectors. End-to-end object detection has emerged as a paradigm shift from traditional pipelines, offering streamlined architectures [53]. DETR [4] introduces the transformer architecture and adopts Hungarian loss to achieve one-to-one matching prediction, thereby eliminating hand-crafted components and post-processing. Since then, various DETR variants have been proposed to enhance its performance and efficiency [42, 67, 56, 30, 36, 28, 5, 77, 82]. Deformable-DETR [81] leverages multi-scale deformable attention module to accelerate the convergence speed. DINO [73] integrates contrastive denoising, mix query selection, and look forward twice scheme into DETRs. RT-DETR [78] further designs the efficient hybrid encoder and proposes the uncertainty-minimal query selection to improve both the accuracy and latency. Another line to achieve end-to-end object detection is based CNN detectors. Learnable NMS [24] and relation networks [26] present another network to remove duplicated predictions for detectors. One Net [55] and De FCN [66] propose oneto-one matching strategies to enable end-to-end object detection with fully convolutional networks. FCOSpss [80] introduces a positive sample selector to choose the optimal sample for prediction.

3 Methodology

3.1 Consistent Dual Assignments for NMS-free Training

During training, YOLOs [21, 65, 29, 70] usually leverage TAL [15] to allocate multiple positive samples for each instance. The adoption of one-to-many assignment yields plentiful supervisory signals, facilitating the optimization and achieving superior performance. However, it necessitates YOLOs to rely on the NMS post-processing, which causes the suboptimal inference efficiency for deployment. While previous works [55, 66, 80, 6] explore one-to-one matching to suppress the redundant predictions, they usually introduce additional inference overhead or yield suboptimal performance. In this work, we present a NMS-free training strategy for YOLOs with dual label assignments and consistent matching metric, achieving both high efficiency and competitive performance.

Dual label assignments. Unlike one-to-many assignment, one-to-one matching assigns only one prediction to each ground truth, avoiding the NMS post-processing. However, it leads to weak supervision, which causes suboptimal accuracy and convergence speed [82]. Fortunately, this deficiency can be compensated for by the one-to-many assignment [6]. To achieve this, we introduce

Backbone PAN

Classification

One-to-many Head Dual Label Assignments Consistent Match. Metric

𝑚= 𝑠 𝑝𝛼 Io U 𝑏, 𝑏

Classification

One-to-one Head

Figure 2: (a) Consistent dual assignments for NMS-free training. (b) Frequency of one-to-one assignments in Top-1/5/10 of one-to-many results for YOLOv8-S which employs αo2m=0.5 and βo2m=6 by default [21]. For consistency, αo2o=0.5; βo2o=6. For inconsistency, αo2o=0.5; βo2o =2.

dual label assignments for YOLOs to combine the best of both strategies. Specifically, as shown in Fig. 2.(a), we incorporate another one-to-one head for YOLOs. It retains the identical structure and adopts the same optimization objectives as the original one-to-many branch but leverages the one-to-one matching to obtain label assignments. During training, two heads are jointly optimized with the model, allowing the backbone and neck to enjoy the rich supervision provided by the oneto-many assignment. During inference, we discard the one-to-many head and utilize the one-to-one head to make predictions. This enables YOLOs for the end-to-end deployment without incurring any additional inference cost. Besides, in the one-to-one matching, we adopt the top one selection, which achieves the same performance as Hungarian matching [4] with less extra training time.

Consistent matching metric. During assignments, both one-to-one and one-to-many approaches leverage a metric to quantitatively assess the level of concordance between predictions and instances. To achieve prediction aware matching for both branches, we employ a uniform matching metric, i.e., m(α, β) = s pα Io U(ˆb, b)β, (1) where p is the classification score, ˆb and b denote the bounding box of prediction and instance, respectively. s represents the spatial prior indicating whether the anchor point of prediction is within the instance [21, 65, 29, 70]. α and β are two important hyperparameters that balance the impact of the semantic prediction task and the location regression task. We denote the one-to-many and one-to-one metrics as mo2m=m(αo2m, βo2m) and mo2o=m(αo2o, βo2o), respectively. These metrics influence the label assignments and supervision information for the two heads.

In dual label assignments, the one-to-many branch provides much richer supervisory signals than one-to-one branch. Intuitively, if we can harmonize the supervision of the one-to-one head with that of one-to-many head, we can optimize the one-to-one head towards the direction of one-to-many head s optimization. As a result, the one-to-one head can provide improved quality of samples during inference, leading to better performance. To this end, we first analyze the supervision gap between the two heads. Due to the randomness during training, we initiate our examination in the beginning with two heads initialized with the same values and producing the same predictions, i.e., one-to-one head and one-to-many head generate the same p and Io U for each prediction-instance pair. We note that the regression targets of two branches do not conflict, as matched predictions share the same targets and unmatched predictions are ignored. The supervision gap thus lies in the different classification targets. Given an instance, we denote its largest Io U with predictions as u , and the largest one-to-many and one-to-one matching scores as m o2m and m o2o, respectively. Suppose that one-to-many branch yields the positive samples Ωand one-to-one branch selects i-th prediction with the metric mo2o,i=m o2o, we can then derive the classification target to2m,j=u mo2m,j

m o2m u for j Ωand to2o,i=u mo2o,i

for task aligned loss as in [21, 65, 29, 70, 15]. The supervision gap between two branches can thus be derived by the 1-Wasserstein distance [46] of different classification objectives, i.e., A = to2o,i I(i Ω)to2m,i + X

k Ω\{i} to2m,k, (2)

We can observe that the gap decreases as to2m,i increases, i.e., i ranks higher within Ω. It reaches the minimum when to2m,i=u , i.e., i is the best positive sample in Ω, as shown in Fig. 2.(a). To achieve this, we present the consistent matching metric, i.e., αo2o=r αo2m and βo2o=r βo2m, which implies mo2o=mr o2m. Therefore, the best positive sample for one-to-many head is also the best for one-to-one head. Consequently, both heads can be optimized consistently and harmoniously. For simplicity, we

take r=1, by default, i.e., αo2o=αo2m and βo2o=βo2m. To verify the improved supervision alignment, we count the number of one-to-one matching pairs within the top-1 / 5 / 10 of the one-to-many results after training. As shown in Fig. 2.(b), the alignment is improved under the consistent matching metric. For a more comprehensive understanding of the mathematical proof, please refer to the appendix.

Discussion with other counter-parts. Similarly, previous works [28, 5, 77, 54, 6, 82, 45] explore the different assignments to accelerate the training convergence and improve the performance for different networks. For example, H-DETR [28], Group-DETR [5], and MS-DETR [77] introduce one-to-many matching in conjunction with the original one-to-one matching by hybrid or multiple group label assignments, to improve upon DETR. Differently, to achieve the one-to-many matching, they usually introduce extra queries or repeat ground truths for bipartite matching, or select top several queries from the matching scores, while we adopt the prediction aware assignment that incorporates the spatial prior. Besides, LRANet [54] employs the dense assignment and sparse assignment branches for training, which all belong to the one-to-many assignment, while we adopt the one-to-many and one-to-one branches. DEYO [45, 43, 44] investigates the step-by-step training with one-to-many matching in the first stage for convolutional encoder and one-to-one matching in the second stage for transformer decoder, while we avoid the transformer decoder for end-to-end inference. Compared with works [6, 80] which incorporate dual assignments for CNN-based detectors, we further analyze the supervision gap between the two heads and present the consistent matching metric for YOLOs to reduce the theoretical supervision gap. It improves performance through better supervision alignment and eliminates the need for hyper-parameter tuning.

3.2 Holistic Efficiency-Accuracy Driven Model Design

In addition to the post-processing, the model architectures of YOLOs also pose great challenges to the efficiency-accuracy trade-offs [50, 8, 29]. Although previous works explore various design strategies, the comprehensive inspection for various components in YOLOs is still lacking. Consequently, the model architecture exhibits non-negligible computational redundancy and constrained capability, which impedes its potential for achieving high efficiency and performance. Here, we aim to holistically perform model designs for YOLOs from both efficiency and accuracy perspectives.

Efficiency driven model design. The components in YOLO consist of the stem, downsampling layers, stages with basic building blocks, and the head. The stem incurs few computational cost and we thus perform efficiency driven model design for other three parts.

(1) Lightweight classification head. The classification and regression heads usually share the same architecture in YOLOs. However, they exhibit notable disparities in computational overhead. For example, the FLOPs and parameter count of the classification head (5.95G/1.51M) are 2.5 and 2.4 those of the regression head (2.34G/0.64M) in YOLOv8-S, respectively. However, after analyzing the impact of classification error and the regression error (seeing Tab. 6), we find that the regression head undertakes more significance for the performance of YOLOs. Consequently, we can reduce the overhead of classification head without worrying about hurting the performance greatly. Therefore, we simply adopt a lightweight architecture for the classification head, which consists of two depthwise separable convolutions [25, 9] with the kernel size of 3 3 followed by a 1 1 convolution.

(2) Spatial-channel decoupled downsampling. YOLOs typically leverage regular 3 3 standard convolutions with stride of 2, achieving spatial downsampling (from H W to H

2 ) and channel transformation (from C to 2C) simultaneously. This introduces non-negligible computational cost of O( 9

2HWC2) and parameter count of O(18C2). Instead, we propose to decouple the spatial reduction and channel increase operations, enabling more efficient downsampling. Specifically, we firstly leverage the pointwise convolution to modulate the channel dimension and then utilize the depthwise convolution to perform spatial downsampling. This reduces the computational cost to O(2HWC2 + 9 2HWC) and the parameter count to O(2C2 +18C). Meanwhile, it maximizes information retention during downsampling, leading to competitive performance with latency reduction.

(3) Rank-guided block design. YOLOs usually employ the same basic building block for all stages [29, 65], e.g., the bottleneck block in YOLOv8 [21]. To thoroughly examine such homogeneous design for YOLOs, we utilize the intrinsic rank [33, 16] to analyze the redundancy2 of each stage. Specifically, we calculate the numerical rank of the last convolution in the last basic block in each stage, which counts the number of singular values larger than a threshold. Fig. 3.(a) presents the results of

2A lower rank implies greater redundancy, while a higher rank signifies more condensed information.

(a) (b) (c)

Figure 3: (a) The intrinsic ranks across stages and models in YOLOv8. The stage in the backbone and neck is numbered in the order of model forward process. The numerical rank r is normalized to r/Co for y-axis and its threshold is set to λmax/2, by default, where Co denotes the number of output channels and λmax is the largest singular value. It can be observed that deep stages and large models exhibit lower intrinsic rank values. (b) The compact inverted block (CIB). (c) The partial self-attention module (PSA).

YOLOv8, indicating that deep stages and large models are prone to exhibit more redundancy. This observation suggests that simply applying the same block design for all stages is suboptimal for the best capacity-efficiency trade-off. To tackle this, we propose a rank-guided block design scheme which aims to decrease the complexity of stages that are shown to be redundant using compact architecture design. We first present a compact inverted block (CIB) structure, which adopts the cheap depthwise convolutions for spatial mixing and cost-effective pointwise convolutions for channel mixing, as shown in Fig. 3.(b). It can serve as the efficient basic building block, e.g., embedded in the ELAN structure [64, 21] (Fig. 3.(b)). Then, we advocate a rank-guided block allocation strategy to achieve the best efficiency while maintaining competitive capacity. Specifically, given a model, we sort its all stages based on their intrinsic ranks in ascending order. We further inspect the performance variation of replacing the basic block in the leading stage with CIB. If there is no performance degradation compared with the given model, we proceed with the replacement of the next stage and halt the process otherwise. Consequently, we can implement adaptive compact block designs across stages and model scales, achieving higher efficiency without compromising performance. Due to the page limit, we provide the details of the algorithm in the appendix.

Accuracy driven model design. We further explore the large-kernel convolution and self-attention for accuracy driven design, aiming to boost the performance under minimal cost.

(1) Large-kernel convolution. Employing large-kernel depthwise convolution is an effective way to enlarge the receptive field and enhance the model s capability [10, 40, 39]. However, simply leveraging them in all stages may introduce contamination in shallow features used for detecting small objects, while also introducing significant I/O overhead and latency in high-resolution stages [8]. Therefore, we propose to leverage the large-kernel depthwise convolutions in CIB within the deep stages. Specifically, we increase the kernel size of the second 3 3 depthwise convolution in the CIB to 7 7, following [39]. Additionally, we employ the structural reparameterization technique [11, 10, 59] to bring another 3 3 depthwise convolution branch to alleviate the optimization issue without inference overhead. Furthermore, as the model size increases, its receptive field naturally expands, with the benefit of using large-kernel convolutions diminishing. Therefore, we only adopt large-kernel convolution for small model scales.

(2) Partial self-attention (PSA). Self-attention [58] is widely employed in various visual tasks due to its remarkable global modeling capability [38, 14, 76]. However, it exhibits high computational complexity and memory footprint. To address this, in light of the prevalent attention head redundancy [69], we present an efficient partial self-attention (PSA) module design, as shown in Fig. 3.(c). Specifically, we evenly partition the features across channels into two parts after the 1 1 convolution. We only feed one part into the NPSA blocks comprised of multi-head self-attention module (MHSA) and feed-forward network (FFN). Two parts are then concatenated and fused by a 1 1 convolution. Besides, we follow [22] to assign the dimensions of the query and key to half of that of the value in MHSA and replace the Layer Norm [1] with Batch Norm [27] for fast inference. Furthermore, PSA is only placed after the Stage 4 with the lowest resolution, avoiding the excessive overhead from the

Table 1: Comparisons with state-of-the-arts. Latency is measured using official pre-trained models. Latencyf denotes the latency in the forward process of model without post-processing. means the results of YOLOv10 with the original one-to-many training using NMS. All results below are without the additional advanced training techniques like knowledge distillation or PGI for fair comparisons.

Model #Param.(M) FLOPs(G) APval(%) Latency(ms) Latencyf(ms) YOLOv6-3.0-N [29] 4.7 11.4 37.0 2.69 1.76 Gold-YOLO-N [60] 5.6 12.1 39.6 2.92 1.82 YOLOv8-N [21] 3.2 8.7 37.3 6.16 1.77 YOLOv10-N (Ours) 2.3 6.7 38.5 / 39.5 1.84 1.79 YOLOv6-3.0-S [29] 18.5 45.3 44.3 3.42 2.35 Gold-YOLO-S [60] 21.5 46.0 45.4 3.82 2.73 YOLO-MS-XS [8] 4.5 17.4 43.4 8.23 2.80 YOLO-MS-S [8] 8.1 31.2 46.2 10.12 4.83 YOLOv8-S [21] 11.2 28.6 44.9 7.07 2.33 YOLOv9-S [65] 7.1 26.4 46.7 - - RT-DETR-R18 [78] 20.0 60.0 46.5 4.58 4.49 YOLOv10-S (Ours) 7.2 21.6 46.3 / 46.8 2.49 2.39 YOLOv6-3.0-M [29] 34.9 85.8 49.1 5.63 4.56 Gold-YOLO-M [60] 41.3 87.5 49.8 6.38 5.45 YOLO-MS [8] 22.2 80.2 51.0 12.41 7.30 YOLOv8-M [21] 25.9 78.9 50.6 9.50 5.09 YOLOv9-M [65] 20.0 76.3 51.1 - - RT-DETR-R34 [78] 31.0 92.0 48.9 6.32 6.21 RT-DETR-R50m [78] 36.0 100.0 51.3 6.90 6.84 YOLOv10-M (Ours) 15.4 59.1 51.1 / 51.3 4.74 4.63 YOLOv6-3.0-L [29] 59.6 150.7 51.8 9.02 7.90 Gold-YOLO-L [60] 75.1 151.7 51.8 10.65 9.78 YOLOv9-C [65] 25.3 102.1 52.5 10.57 6.13 YOLOv10-B (Ours) 19.1 92.0 52.5 / 52.7 5.74 5.67 YOLOv8-L [21] 43.7 165.2 52.9 12.39 8.06 RT-DETR-R50 [78] 42.0 136.0 53.1 9.20 9.07 YOLOv10-L (Ours) 24.4 120.3 53.2 / 53.4 7.28 7.21 YOLOv8-X [21] 68.2 257.8 53.9 16.86 12.83 RT-DETR-R101 [78] 76.0 259.0 54.3 13.71 13.58 YOLOv10-X (Ours) 29.5 160.4 54.4 / 54.4 10.70 10.60

quadratic computational complexity of self-attention. In this way, the global representation learning ability can be incorporated into YOLOs with low computational costs, which well enhances the model s capability and leads to improved performance.

4 Experiments

4.1 Implementation Details

We select YOLOv8 [21] as our baseline model, due to its commendable latency-accuracy balance and its availability in various model sizes. We employ the consistent dual assignments for NMS-free training and perform holistic efficiency-accuracy driven model design based on it, which brings our YOLOv10 models. YOLOv10 has the same variants as YOLOv8, i.e., N / S / M / L / X. Besides, we derive a new variant YOLOv10-B, by simply increasing the width scale factor of YOLOv10-M. We verify the proposed detector on COCO [35] under the same training-from-scratch setting [21, 65, 62]. Moreover, the latencies of all models are tested on T4 GPU with Tensor RT FP16, following [78].

4.2 Comparison with state-of-the-arts

As shown in Tab. 1, our YOLOv10 achieves the state-of-the-art performance and end-to-end latency across various model scales. We first compare YOLOv10 with our baseline models, i.e., YOLOv8. On N / S / M / L / X five variants, our YOLOv10 achieves 1.2% / 1.4% / 0.5% / 0.3% / 0.5% AP improvements, with 28% / 36% / 41% / 44% / 57% fewer parameters, 23% / 24% / 25% / 27% / 38% less calculations, and 70% / 65% / 50% / 41% / 37% lower latencies. Compared with other YOLOs,

Table 2: Ablation study with YOLOv10-S and YOLOv10-M on COCO.

# Model NMS-free. Efficiency. Accuracy. #Param.(M) FLOPs(G) APval(%) Latency(ms) 1

11.2 28.6 44.9 7.07 2 11.2 28.6 44.3 2.44 3 6.2 20.8 44.5 2.31 4 7.2 21.6 46.3 2.49 5

25.9 78.9 50.6 9.50 6 25.9 78.9 50.3 5.22 7 14.1 58.1 50.4 4.57 8 15.4 59.1 51.1 4.74

Table 3: Dual assign.

o2m o2o AP Latency

44.9 7.07 43.4 2.44 44.3 2.44

Table 4: Matching metric.

αo2o βo2o APval αo2o βo2o APval

0.5 2.0 42.7 0.25 3.0 44.3 0.5 4.0 44.2 0.25 6.0 43.5 0.5 6.0 44.3 1.0 6.0 43.9 0.5 8.0 44.0 1.0 12.0 44.3

Table 5: Efficiency. for YOLOv10-S/M.

# Model #Param FLOPs APval Latency 1 base. 11.2/25.9 28.6/78.9 44.3/50.3 2.44/5.22 2 +cls. 9.9/23.2 23.5/67.7 44.2/50.2 2.39/5.07 3 +downs. 8.0/19.7 22.2/65.0 44.4/50.4 2.36/4.97 4 +block. 6.2/14.1 20.8/58.1 44.5/50.4 2.31/4.57

YOLOv10 also exhibits superior trade-offs between accuracy and computational cost. Specifically, for lightweight and small models, YOLOv10-N / S outperforms YOLOv6-3.0-N / S by 1.5 AP and 2.0 AP, with 51% / 61% fewer parameters and 41% / 52% less computations, respectively. For medium models, compared with YOLOv9-C / YOLO-MS, YOLOv10-B / M enjoys the 46% / 62% latency reduction under the same or better performance, respectively. For large models, compared with Gold-YOLO-L, our YOLOv10-L shows 68% fewer parameters and 32% lower latency, along with a significant improvement of 1.4% AP. Furthermore, compared with RT-DETR, YOLOv10 obtains significant performance and latency improvements. Notably, YOLOv10-S / X achieves 1.8 and 1.3 faster inference speed than RT-DETR-R18 / R101, respectively, under the similar performance. These results well demonstrate the superiority of YOLOv10 as the real-time end-to-end detector.

We also compare YOLOv10 with other YOLOs using the original one-to-many training approach. We consider the performance and the latency of model forward process (Latencyf) in this situation, following [62, 21, 60]. As shown in Tab. 1, YOLOv10 also exhibits the state-of-the-art performance and efficiency across different model scales, indicating the effectiveness of our architectural designs.

4.3 Model Analyses

Ablation study. We present the ablation results based on YOLOv10-S and YOLOv10-M in Tab. 2. It can be observed that our NMS-free training with consistent dual assignments significantly reduces the end-to-end latency of YOLOv10-S by 4.63ms, while maintaining competitive performance of 44.3% AP. Moreover, our efficiency driven model design leads to the reduction of 11.8 M parameters and 20.8 GFl OPs, with a considerable latency reduction of 0.65ms for YOLOv10-M, well showing its effectiveness. Furthermore, our accuracy driven model design achieves the notable improvements of 1.8 AP and 0.7 AP for YOLOv10-S and YOLOv10-M, alone with only 0.18ms and 0.17ms latency overhead, respectively, which well demonstrates its superiority.

Analyses for NMS-free training. Dual label assignments. We present dual label assignments for NMS-free YOLOs, which can bring both rich supervision of one-to-many (o2m) branch during training and high efficiency of one-to-one (o2o) branch during inference. We verify its benefit based on YOLOv8-S, i.e., #1 in Tab. 2. Specifically, we introduce baselines for training with only o2m branch and only o2o branch, respectively. As shown in Tab. 3, our dual label assignments achieve the best AP-latency trade-off. Consistent matching metric. We introduce consistent matching metric to make the one-to-one head more harmonious with the one-to-many head. We verify its benefit based on YOLOv8-S, i.e., #1 in Tab. 2, under different αo2o and βo2o. As shown in Tab. 4, the proposed consistent matching metric, i.e., αo2o=r αo2m and βo2o=r βo2m, can achieve the optimal performance, where αo2m=0.5 and βo2m=6.0 in the one-to-many head [21]. Such an improvement can be attributed to the reduction of the supervision gap (Eq. (2)), which provides improved supervision alignment between two branches. Moreover, the proposed consistent matching metric eliminates the need for exhaustive hyper-parameter tuning, which is appealing in practical scenarios.

Figure 4: The average cosine similarity of each anchor point s extracted features with all others.

Table 6: cls. results.

base. +cls. APval 44.3 44.2 APval w/o c 59.9 59.9 APval w/o r 64.5 64.2

Table 7: Results of d.s.

Model APval Latency base. 43.7 2.33 ours 44.4 2.36

Table 8: Results of CIB.

Model APval Latency IRB 43.7 2.30 IRB-DW 44.2 2.30 ours 44.5 2.31

Table 9: Rank-guided.

Stages with CIB APval

empty 44.4 8 44.5 8,4, 44.5 8,4,7 44.3

Performance gap compared with one-to-many training. Although achieving superior end-to-end performance under NMS-free training, we observe that there still exists the performance gap compared with the original one-to-many training using NMS, as shown in Tab. 3 and Tab. 1. Besides, we note that the gap diminishes as the model size increases. Therefore, we reasonably concludes that such a gap can be attributed to the limitations in the model capability. Notably, unlike the original one-to-many training using NMS, the NMS-free training necessitates more discriminative features for one-to-one matching. In the case of the YOLOv10-N model, its limited capacity results in extracted features that lack sufficient discriminability, leading to a more notable performance gap of 1.0% AP. In contrast, the YOLOv10-X model, which possesses stronger capability and more discriminative features, shows no performance gap between two training strategies. In Fig. 4, we visualize the average cosine similarity of each anchor point s extracted features with those of all other anchor points on the COCO val set. We observe that as the model size increases, the feature similarity between anchor points exhibits a downward trend, which benefits the one-to-one matching. Based on this insight, we will explore approaches to further reduce the gap and achieve higher end-to-end performance in the future work. Analyses for efficiency driven model design. We conduct experiments to gradually incorporate the efficiency driven design elements based on YOLOv10-S/M. Our baseline is the YOLOv10-S/M model without efficiency-accuracy driven model design, i.e., #2/#6 in Tab. 2. As shown in Tab. 5, each design component, including lightweight classification head, spatial-channel decoupled downsampling, and rank-guided block design, contributes to the reduction of parameters count, FLOPs, and latency. Importantly, these improvements are achieved while maintaining competitive performance. Lightweight classification head. We analyze the impact of category and localization errors of predictions on the performance, based on the YOLOv10-S of #1 and #2 in Tab. 5, like [7]. Specifically, we match the predictions to the instances by the one-to-one assignment. Then, we substitute the predicted category score with instance labels, resulting in APval w/o c with no classification errors. Similarly, we replace the predicted locations with those of instances, yielding APval w/o r with no regression errors. As shown in Tab. 6, APval w/o r is much higher than APval w/o c, revealing that eliminating the regression errors achieves greater improvement. The performance bottleneck thus lies more in the regression task. Therefore, adopting the lightweight classification head can allow higher efficiency without compromising the performance. Spatial-channel decoupled downsampling. We decouple the downsampling operations for efficiency, where the channel dimensions are first increased by pointwise convolution (PW) and the resolution is then reduced by depthwise convolution (DW) for maximal information retention. We compare it with the baseline way of spatial reduction by DW followed by channel modulation by PW, based on the YOLOv10-S of #3 in Tab. 5. As shown in Tab. 7, our downsampling strategy achieves the 0.7% AP improvement by enjoying less information loss during downsampling. Compact inverted block (CIB). We introduce CIB as the compact basic building block. We verify its effectiveness based on the YOLOv10-S of #4 in the Tab. 5. Specifically, we introduce the inverted residual block [51] (IRB) as the baseline, which achieves the suboptimal 43.7% AP, as shown in Tab. 8. We then append a 3 3 depthwise convolution (DW) after it, denoted as IRB-DW , which

Table 10: Accuracy. for S/M.

# Model APval Latency 1 base. 44.5/50.4 2.31/4.57 2 +L.k. 44.9/- 2.34/- 3 +PSA 46.3/51.1 2.49/4.74

Table 11: L.k. results.

Model APval Latency k.s.=5 44.7 2.32 k.s.=7 44.9 2.34 k.s.=9 44.9 2.37 w/o rep. 44.8 2.34

Table 12: L.k. usage.

w/o L.k. w/ L.k. N 36.3 36.6 S 44.5 44.9 M 50.4 50.4

Table 13: PSA results.

Model APval Latency PSA 46.3 2.49 Trans. 46.0 2.54 NPSA = 1 46.3 2.49 NPSA = 2 46.5 2.59

brings 0.5% AP improvement. Compared with IRB-DW , our CIB further achieves 0.3% AP improvement by prepending another DW with minimal overhead, indicating its superiority. Rank-guided block design. We introduce the rank-guided block design to adaptively integrate compact block design for improving the model efficiency. We verify its benefit based on the YOLOv10-S of #3 in the Tab. 5. The stages sorted in ascending order based on the intrinsic ranks are Stage 8-4-7-3-5-1-6-2, like in Fig. 3.(a). As shown in Tab. 9, when gradually replacing the bottleneck block in each stage with the efficient CIB, we observe the performance degradation starting from Stage 7. In the Stage 8 and 4 with lower intrinsic ranks and more redundancy, we can thus adopt the efficient block design without compromising the performance. These results indicate that rank-guided block design can serve as an effective strategy for higher model efficiency. Analyses for accuracy driven model design. We present the results of gradually integrating the accuracy driven design elements based on YOLOv10-S/M. Our baseline is the YOLOv10-S/M model after incorporating efficiency driven design, i.e., #3/#7 in Tab. 2. As shown in Tab. 10, the adoption of large-kernel convolution and PSA module leads to the considerable performance improvements of 0.4% AP and 1.4% AP for YOLOv10-S under minimal latency increase of 0.03ms and 0.15ms, respectively. Note that large-kernel convolution is not employed for YOLOv10-M (see Tab. 12). Large-kernel convolution. We first investigate the effect of different kernel sizes based on the YOLOv10-S of #2 in Tab. 10. As shown in Tab. 11, the performance improves as the kernel size increases and stagnates around the kernel size of 7 7, indicating the benefit of large perception field. Besides, removing the reparameterization branch during training achieves 0.1% AP degradation, showing its effectiveness for optimization. Moreover, we inspect the benefit of large-kernel convolution across model scales based on YOLOv10-N / S / M. As shown in Tab. 12, it brings no improvements for large models, i.e., YOLOv10-M, due to its inherent extensive receptive field. We thus only adopt large-kernel convolutions for small models, i.e., YOLOv10-N / S. Partial self-attention (PSA). We introduce PSA to enhance the performance by incorporating the global modeling ability under minimal cost. We first verify its effectiveness based on the YOLOv10S of #3 in Tab. 10. Specifically, we introduce the transformer block, i.e., MHSA followed by FFN, as the baseline, denoted as Trans. . As shown in Tab. 13, compared with it, PSA brings 0.3% AP improvement with 0.05ms latency reduction. The performance enhancement may be attributed to the alleviation of optimization problem [68, 10] in self-attention, by mitigating the redundancy in attention heads. Moreover, we investigate the impact of different NPSA. As shown in Tab. 13, increasing NPSA to 2 obtains 0.2% AP improvement but with 0.1ms latency overhead. Therefore, we set NPSA to 1, by default, to enhance the model capability while maintaining high efficiency.

5 Conclusion

In this paper, we target both the post-processing and model architecture throughout the detection pipeline of YOLOs. For the post-processing, we propose the consistent dual assignments for NMSfree training, achieving efficient end-to-end detection. For the model architecture, we introduce the holistic efficiency-accuracy driven model design strategy, improving the performance-efficiency tradeoffs. These bring our YOLOv10, a new real-time end-to-end object detector. Extensive experiments show that YOLOv10 achieves the state-of-the-art performance and latency compared with other advanced detectors, well demonstrating its superiority.

6 Acknowledgments

This work was supported by National Natural Science Foundation of China (Nos. 61925107, 62271281) and Beijing Natural Science Foundation (No. L223023).

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ar Xiv preprint ar Xiv:1607.06450, 2016.

[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection, 2020.

[3] Daniel Bogdoll, Maximilian Nitsche, and J Marius Zöllner. Anomaly detection in autonomous driving: A survey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4488 4499, 2022.

[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213 229. Springer, 2020.

[5] Qiang Chen, Xiaokang Chen, Jian Wang, Shan Zhang, Kun Yao, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise oneto-many assignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6633 6642, 2023.

[6] Yiqun Chen, Qiang Chen, Qinghao Hu, and Jian Cheng. Date: Dual assignment for end-to-end fully convolutional object detection. ar Xiv preprint ar Xiv:2211.13859, 2022.

[7] Yiqun Chen, Qiang Chen, Peize Sun, Shoufa Chen, Jingdong Wang, and Jian Cheng. Enhancing your trained detrs with box refinement. ar Xiv preprint ar Xiv:2307.11828, 2023.

[8] Yuming Chen, Xinbin Yuan, Ruiqi Wu, Jiabao Wang, Qibin Hou, and Ming-Ming Cheng. Yolo-ms: rethinking multi-scale representation learning for real-time object detection. ar Xiv preprint ar Xiv:2308.05480, 2023.

[9] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251 1258, 2017.

[10] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963 11975, 2022.

[11] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13733 13742, 2021.

[12] Douglas Henke Dos Reis, Daniel Welfer, Marco Antonio De Souza Leite Cuadros, and Daniel Fernando Tello Gamarra. Mobile robot navigation using an object recognition software with rgbd images and the yolo algorithm. Applied Artificial Intelligence, 33(14):1290 1305, 2019.

[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569 6578, 2019.

[14] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873 12883, 2021.

[15] Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R Scott, and Weilin Huang. Tood: Task-aligned one-stage object detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3490 3499. IEEE Computer Society, 2021.

[16] Ruili Feng, Kecheng Zheng, Yukun Huang, Deli Zhao, Michael Jordan, and Zheng-Jun Zha. Rank diminishing in deep neural networks. Advances in Neural Information Processing Systems, 35:33054 33065, 2022.

[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. ar Xiv preprint ar Xiv:2107.08430, 2021.

[18] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2918 2928, 2021.

[19] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440 1448, 2015.

[20] Jocher Glenn. Yolov5 release v7.0. https: // github. com/ ultralytics/ yolov5/ tree/ v7. 0 , 2022.

[21] Jocher Glenn. Yolov8. https: // github. com/ ultralytics/ ultralytics/ tree/ main , 2023.

[22] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet s clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259 12269, 2021.

[23] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961 2969, 2017.

[24] Jan Hosang, Rodrigo Benenson, and Bernt Schiele. Learning non-maximum suppression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4507 4515, 2017.

[25] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. ar Xiv preprint ar Xiv:1704.04861, 2017.

[26] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588 3597, 2018.

[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448 456. pmlr, 2015.

[28] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19702 19712, 2023.

[29] Chuyi Li, Lulu Li, Yifei Geng, Hongliang Jiang, Meng Cheng, Bo Zhang, Zaidan Ke, Xiaoming Xu, and Xiangxiang Chu. Yolov6 v3.0: A full-scale reloading. ar Xiv preprint ar Xiv:2301.05586, 2023.

[30] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619 13627, 2022.

[31] Xiang Li, Wenhai Wang, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss v2: Learning reliable localization quality estimation for dense object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11632 11641, 2021.

[32] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 33:21002 21012, 2020.

[33] Ming Lin, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Neural architecture design for gpu-efficient networks. ar Xiv preprint ar Xiv:2006.14090, 2020.

[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980 2988, 2017.

[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740 755. Springer, 2014.

[36] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ar Xiv preprint ar Xiv:2201.12329, 2022.

[37] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759 8768, 2018.

[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012 10022, 2021.

[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976 11986, 2022.

[40] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016.

[41] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. ar Xiv preprint ar Xiv:2212.07784, 2022.

[42] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651 3660, 2021.

[43] Haodong Ouyang. Deyov2: Rank feature with greedy matching for end-to-end object detection. ar Xiv preprint ar Xiv:2306.09165, 2023.

[44] Haodong Ouyang. Deyov3: Detr with yolo for real-time object detection. ar Xiv preprint ar Xiv:2309.11851, 2023.

[45] Haodong Ouyang. Deyo: Detr with yolo for end-to-end object detection. ar Xiv preprint ar Xiv:2402.16370, 2024.

[46] Victor M Panaretos and Yoav Zemel. Statistical aspects of wasserstein distances. Annual review of statistics and its application, 6:405 431, 2019.

[47] Joseph Redmon. Darknet: Open source neural networks in c. http://pjreddie.com/ darknet/, 2013 2016.

[48] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.

[49] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.

[50] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018.

[51] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510 4520, 2018.

[52] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430 8439, 2019.

[53] Russell Stewart, Mykhaylo Andriluka, and Andrew Y Ng. End-to-end people detection in crowded scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2325 2333, 2016.

[54] Yuchen Su, Zhineng Chen, Zhiwen Shao, Yuning Du, Zhilong Ji, Jinfeng Bai, Yong Zhou, and Yu-Gang Jiang. Lranet: Towards accurate and efficient scene text detection with low-rank approximation network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4979 4987, 2024.

[55] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934 9944. PMLR, 2021.

[56] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454 14463, 2021.

[57] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: A simple and strong anchor-free object detector. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):1922 1933, 2020.

[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.

[59] Ao Wang, Hui Chen, Zijia Lin, Hengjun Pu, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. ar Xiv preprint ar Xiv:2307.09283, 2023.

[60] Chengcheng Wang, Wei He, Ying Nie, Jianyuan Guo, Chuanjian Liu, Yunhe Wang, and Kai Han. Gold-yolo: Efficient object detector via gather-and-distribute mechanism. Advances in Neural Information Processing Systems, 36, 2024.

[61] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Scaled-yolov4: Scaling cross stage partial network. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13029 13038, 2021.

[62] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-offreebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464 7475, 2023.

[63] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 390 391, 2020.

[64] Chien-Yao Wang, Hong-Yuan Mark Liao, and I-Hau Yeh. Designing network design strategies through gradient path analysis. ar Xiv preprint ar Xiv:2211.04800, 2022.

[65] Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. ar Xiv preprint ar Xiv:2402.13616, 2024.

[66] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15849 15858, 2021.

[67] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 2567 2575, 2022.

[68] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22 31, 2021.

[69] Haiyang Xu, Zhichao Zhou, Dongliang He, Fu Li, and Jingdong Wang. Vision transformer with attention map hallucination and ffn compaction. ar Xiv preprint ar Xiv:2306.10875, 2023.

[70] Shangliang Xu, Xinxin Wang, Wenyu Lv, Qinyao Chang, Cheng Cui, Kaipeng Deng, Guanzhong Wang, Qingqing Dang, Shengyu Wei, Yuning Du, et al. Pp-yoloe: An evolved version of yolo. ar Xiv preprint ar Xiv:2203.16250, 2022.

[71] Xianzhe Xu, Yiqi Jiang, Weihua Chen, Yilun Huang, Yuan Zhang, and Xiuyu Sun. Damo-yolo: A report on real-time object detection design. ar Xiv preprint ar Xiv:2211.15444, 2022.

[72] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659 675. Springer, 2022.

[73] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ar Xiv preprint ar Xiv:2203.03605, 2022.

[74] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. ar Xiv preprint ar Xiv:1710.09412, 2017.

[75] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759 9768, 2020.

[76] Wenqiang Zhang, Zilong Huang, Guozhong Luo, Tao Chen, Xinggang Wang, Wenyu Liu, Gang Yu, and Chunhua Shen. Topformer: Token pyramid transformer for mobile semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12083 12093, 2022.

[77] Chuyang Zhao, Yifan Sun, Wenhao Wang, Qiang Chen, Errui Ding, Yi Yang, and Jingdong Wang. Ms-detr: Efficient detr training with mixed supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17027 17036, 2024.

[78] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. ar Xiv preprint ar Xiv:2304.08069, 2023.

[79] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993 13000, 2020.

[80] Qiang Zhou and Chaohui Yu. Object detection made simpler by eliminating heuristic nms. IEEE Transactions on Multimedia, 2023.

[81] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. ar Xiv preprint ar Xiv:2010.04159, 2020.

[82] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6748 6758, 2023.

A.1 Implementation Details

Following [21, 62, 65], all YOLOv10 models are trained from scratch using the SGD optimizer for 500 epochs. The SGD momentum and weight decay are set to 0.937 and 5 10 4, respectively. The initial learning rate is 1 10 2 and it decays linearly to 1 10 4. For data augmentation, we adopt the Mosaic [2, 20], Mixup [74] and copy-paste augmentation [18], etc., like [21, 65]. Tab. 14 presents the detailed hyper-parameters. All models are trained on 8 NVIDIA 3090 GPUs. Besides, we increase the width scale factor of YOLOv10-M to 1.0 to obtain YOLOv10-B. For PSA, we employ it after the SPPF module [21] and adopt the expansion factor of 2 for FFN. For CIB, we also adopt the expansion ratio of 2 for the inverted bottleneck block structure. Following [65, 62], we report the standard mean average precision (AP) across different object scales and Io U thresholds on the COCO dataset [35].

Moreover, we follow [78] to establish the end-to-end speed benchmark. Since the execution time of NMS is affected by the input, we thus measure the latency on the COCO val set with the batch size of 1, like [78]. We adopt the same NMS hyperparameters used by the detectors during their validation. The Tensor RT efficient NMSPlugin is appended for post-processing and the I/O overhead is omitted. We report the average latency across all images.

Table 14: Hyper-parameters of YOLOv10.

hyper-parameter YOLOv10-N/S/M/B/L/X

epochs 500 optimizer SGD momentum 0.937 weight decay 5 10 4

warm-up epochs 3 warm-up momentum 0.8 warm-up bias learning rate 0.1 initial learning rate 10 2

final learning rate 10 4

learning rate schedule linear decay box loss gain 7.5 class loss gain 0.5 DFL loss gain 1.5 HSV saturation augmentation 0.7 HSV value augmentation 0.4 HSV hue augmentation 0.015 translation augmentation 0.1 scale augmentation 0.5/0.5/0.9/0.9/0.9/0.9 mosaic augmentation 1.0 Mixup augmentation 0.0/0.0/0.1/0.1/0.15/0.15 copy-paste augmentation 0.0/0.0/0.1/0.1/0.3/0.3 close mosaic epochs 10

A.2 Details of Consistent Matching Metric

We provide the detailed derivation of consistent matching metric here.

As mentioned in the paper, we suppose that the one-to-many positive samples is Ωand the one-toone branch selects i-th prediction. We can then leverage the normalized metric [15] to obtain the classification target for task alignment learning [21, 15, 65, 29, 70], i.e., to2m,j = u mo2m,j

for j Ωand to2o,i = u mo2o,i

m o2o = u . We can thus derive the supervision gap between two branches by the 1-Wasserstein distance [46] of the different classification targets, i.e., A = |(1 to2o,i) (1 I(i Ω)to2m,i)| + X

k Ω\{i} |1 (1 to2m,k)|

= |to2o,i I(i Ω)to2m,i| + X

k Ω\{i} to2m,k

= to2o,i I(i Ω)to2m,i + X

k Ω\{i} to2m,k,

where I( ) is the indicator function. We denote the classification targets of the predictions in Ωas {ˆt1, ˆt2, ..., ˆt|Ω|} in descending order, with ˆt1 ˆt2 ... ˆt|Ω|. We can then replace to2o,i with u and obtain: A = u I(i Ω)to2m,i + X

k Ω\{i} to2m,k

k Ωto2m,k 2 I(i Ω)to2m,i

k=1 ˆtk 2 I(i Ω)to2m,i

We further discuss the supervision gap in two scenarios, i.e.,

1. Supposing i Ω, we can obtain:

A = u + X|Ω|

k=1 ˆtk (5)

2. Supposing i Ω, we denote to2m,i = ˆtn and obtain:

A = u + X|Ω|

k=1 ˆtk 2 ˆtn (6)

Due to ˆtn 0, the second case can lead to smaller supervision gap. Besides, we can observe that A decreases as ˆtn increases, indicating that n decreases and the ranking of i within Ωimproves. Due to ˆtn ˆt1, A thus achieves the minimum when ˆtn = ˆt1, i.e., i is the best positive sample in Ωwith mo2m,i = m o2m and to2m,i = u mo2m,i

m o2m = u .

Furthermore, we prove that we can achieve the minimized supervision gap by the consistent matching metric. We suppose αo2m > 0 and βo2m > 0, which are common in [21, 65, 29, 15, 70]. Similarly, we assume αo2o > 0 and βo2o > 0. We can obtain r1 = αo2o αo2m > 0 and r2 = βo2o βo2m > 0, and then derive mo2o by mo2o = s pαo2o Io U(ˆb, b)βo2o

= s pr1 αo2m Io U(ˆb, b)r2 βo2m

= s (pαo2m Io U(ˆb, b)βo2m)r1 Io U(ˆb, b)(r2 r1) βo2m

= mr1 o2m Io U(ˆb, b)(r2 r1) βo2m

To achieve mo2m,i = m o2m and mo2o,i = m o2o, we can make mo2o monotonically increase with mo2m by assigning (r2 r1) = 0, i.e., mo2o = mr1 o2m Io U(ˆb, b)0 βo2m

= mr1 o2m (8)

Supposing r1 = r2 = r, we can thus derive the consistent matching metric, i.e., αo2o = r αo2m and βo2o = r βo2m. By simply taking r = 1, we obtain αo2o = αo2m and βo2o = βo2m.

A.3 Details of Rank-Guided Block Design

We present the details of the algorithm of rank-guided block design in Algo. 1. Besides, to calculate the numerical rank of the convolution, we reshape its weight to the shape of (Co, K2 Ci), where Co and Ci denote the number of output and input channels, and K means the kernel size, respectively.

A.4 Training Cost Analyses

In addition to the inference efficiency analyses, we also investigate the training cost of our YOLOv10 models. We compare with other YOLO variants and measure the training throughput on 8 NVIDIA 3090 GPUs using the official codebases. Tab. 15 presents the comparison results based on the medium model scale. We observe that despite having 500 training epochs, YOLOv10 achieves a high training throughput, making its training cost affordable. We also note that the one-to-many head in the NMS-free training will introduce the extra overhead for YOLOv10. To investigate this, we measure the training cost of YOLOv10 with only the one-to-one head, which is denoted as YOLOv10-o2o . As shown in Tab. 15, YOLOv10-M results in a small increase in the training time over YOLOv10M-o2o , about 18s each epoch, which is affordable. To fairly verify the benefit of the one-to-many head in NMS-free training, we also adopt longer 550 training epochs for YOLOv10-M-o2o , which

Algorithm 1: Rank-guided block design Input: Intrinsic ranks R for all stages S; Original Network Θ; CIB θcib; Output: New network Θ with CIB for certain stages.

2 Θ0 Θ; Θ Θ0;

3 ap0 AP(T(Θ0)) ; // T:training the network; AP:evaluating the AP performance.

4 while S = do

5 st argmins S R;

6 Θt+1 Replace(Θt, θcib, st) ; // Replace the block in Stage st of Θt with CIB θcib.

7 apt+1 AP(T(Θt+1));

8 if apt+1 ap0 then

9 Θ Θt+1; S S \ {st};

11 return Θ ;

14 return Θ ;

Table 15: Training cost analyses on 8 NVIDIA 3090 GPUs.

Model Epoch Speed (epoch/hour) Time (hour)

YOLOv6-3.0-M 300 7.2 41.7 YOLOv8-M 500 18.3 27.3 YOLOv9-M 500 12.3 40.7 Gold-YOLO-M 300 4.7 63.8 YOLO-MS 300 7.1 42.3 YOLOv10-M-o2o 500 18.8 26.7 YOLOv10-M 500 17.2 29.1

Table 16: Latency with NMS.

Model Latency

YOLOv10-N 6.19ms YOLOv10-S 7.15ms YOLOv10-M 9.03ms YOLOv10-B 10.04ms YOLOv10-L 11.52ms YOLOv10-X 14.67ms

leads to a similar training time (29.3 vs. 29.1 hours) but still yields inferior performance (48.9% vs. 51.1% AP) compared with YOLOv10-M.

A.5 More Results on COCO

We measure the latency of YOLOv10 with the original one-to-many training using NMS and report the results on COCO in Tab. 16. Besides, we report the detailed performance of YOLOv10, including APval 50 and APval 75 at different Io U thresholds, as well as APval small, APval medium, and APval large across different scales, in Tab. 17. We also present the comparisons with more lightweight detectors, including DAMO-YOLO [71], YOLOv7 [62], and DEYO [45], in Tab. 18. It shows that our YOLOv10 also achieves superior performance and efficiency trade-offs. Additionally, in experiments, we follow previous works [21, 65] to train the models for 500 epochs. We also conduct experiments to train the models for 300 epochs and present the comparison results with YOLOv6 [29], Gold YOLO [60], and YOLO-MS [8] which adopt 300 epochs, in Tab. 19. We observe that our YOLOv10 also exhibits better performance and inference latency. We also note that despite trained for 500 epochs, YOLOv10 has less training cost compared with these models as presented in Tab. 15.

A.6 Inference Efficiency Comparison on CPU

We present the speed comparison results of YOLOv10 and others on CPU (Intel Xeon Skylake, IBRS) using Open VINO in Fig. 5. We observe that YOLOv10 also shows state-of-the-art trade-offs in terms of performance and efficiency.

A.7 More Analyses for Holistic Efficiency-Accuracy Driven Model Design

We note that reducing the latency of YOLOv10-S (#2 in Tab. 2) is particularly challenging due to its small model scale. However, as shown in Tab. 2, our efficiency driven model design still achieves a 5.3% reduction in latency without compromising performance. This provides substantial support for the further accuracy driven model design. YOLOv10-S achieves a better latency-accuracy trade-off

Table 17: Detailed performance of YOLOv10 on COCO.

Model APval(%) APval 50 (%) APval 75 (%) APval small(%) APval medium(%) APval large(%)

YOLOv10-N 38.5 53.8 41.7 18.9 42.4 54.6 YOLOv10-S 46.3 63.0 50.4 26.8 51.0 63.8 YOLOv10-M 51.1 68.1 55.8 33.8 56.5 67.0 YOLOv10-B 52.5 69.6 57.2 35.1 57.8 68.5 YOLOv10-L 53.2 70.1 58.1 35.8 58.5 69.4 YOLOv10-X 54.4 71.3 59.3 37.0 59.8 70.9

Table 18: Comparisons with more lightweight detectors.

Model #Param.(M) FLOPs(G) APval(%) Latency(ms)

DEYO-tiny [45] 4.0 8.0 37.6 2.01 YOLOv10-N 2.3 6.7 38.5 1.84 DAMO-YOLO-T [71] 8.5 18.1 42.0 2.21 DAMO-YOLO-S [71] 16.3 37.8 46.0 3.18 DEYO-S [45] 14.0 26.0 45.8 3.34 YOLOv10-S 7.2 21.6 46.3 2.49 DAMO-YOLO-M [71] 28.2 61.8 49.2 4.57 DAMO-YOLO-L [71] 42.1 97.3 50.8 6.48 DEYO-M [45] 33.0 78.0 50.7 7.14 YOLOv10-M 15.4 59.1 51.1 4.74 YOLOv7 [62] 36.9 104.7 51.2 17.03 YOLOv10-B 19.1 92.0 52.5 5.74 YOLOv7-X [62] 71.3 189.9 52.9 21.45 DEYO-L [45] 51.0 155.0 52.7 10.00 YOLOv10-L 24.4 120.3 53.2 7.28 DEYO-X [45] 78.0 242.0 53.7 15.38 YOLOv10-X 29.5 160.4 54.4 10.70

with our holistic efficiency-accuracy driven model design, showing a 2.0% AP improvement with only 0.05ms latency overhead. Besides, for YOLOv10-M (#6 in Tab. 2), which has a larger model scale and more redundancy, our efficiency driven model design results in a considerable 12.5% latency reduction, as shown in Tab. 2. When combined with accuracy driven model design, we observe a notable 0.8% AP improvement for YOLOv10-M, along with a favorable latency reduction of 0.48ms. These results well demonstrate the effectiveness of our design strategy across different model scales.

Table 19: Performance comparisons under 300 training epochs.

Model #Param.(M) FLOPs(G) APval(%) Latency(ms)

YOLOv6-3.0-N [29] 4.7 11.4 37.0 2.69 Gold-YOLO-N [60] 5.6 12.1 39.6 2.92 YOLOv10-N (Ours) 2.3 6.7 37.7 1.84 YOLOv6-3.0-S [29] 18.5 45.3 44.3 3.42 Gold-YOLO-S [60] 21.5 46.0 45.4 3.82 YOLO-MS-XS [8] 4.5 17.4 43.4 8.23 YOLOv10-S (Ours) 7.2 21.6 45.6 2.49 YOLOv6-3.0-M [29] 34.9 85.8 49.1 5.63 Gold-YOLO-M [60] 41.3 87.5 49.8 6.38 YOLOv10-M (Ours) 15.4 59.1 50.3 4.74 YOLOv6-3.0-L [29] 59.6 150.7 51.8 9.02 Gold-YOLO-L [60] 75.1 151.7 51.8 10.65 YOLO-MS [8] 22.2 80.2 51.0 12.41 YOLOv10-B (Ours) 19.1 92.0 51.6 5.74 YOLOv10-L (Ours) 24.4 120.3 52.4 7.28 YOLOv10-X (Ours) 29.5 160.4 53.6 10.70

100 200 300 400 500 Latency on CPU (ms)

COCO AP (%)

YOLOv6-v3.0 YOLOv8 YOLOv9 YOLO-MS Gold-YOLO RT-DETR YOLOv10 (Ours)

Figure 5: Performance and efficiency comparisons on CPU.

Figure 6: Visualization results under complex and challenging scenarios.

A.8 Visualization Results

Fig. 6 presents the visualization results of our YOLOv10 in the complex and challenging scenarios. It can be observed that YOLOv10 can achieve precise detection under various difficult conditions, such as low light, rotation, etc. It also demonstrates a strong capability in detecting diverse and densely packed objects, such as bottle, cup, and person. These results indicate its superior performance.

A.9 Contribution, Limitation, and Broader Impact

Contribution. In summary, our contributions are three folds as follows:

1. We present a novel consistent dual assignments strategy for NMS-free YOLOs. A dual label assignments way is designed to provide rich supervision by one-to-many branch during training and high efficiency by one-to-one branch during inference. Besides, to ensure the harmonious supervision between two branches, we innovatively propose the consistent matching metric, which can well reduce the theoretical supervision gap and lead to improved performance.

2. We propose a holistic efficiency-accuracy driven model design strategy for the model architecture of YOLOs. We present novel lightweight classification head, spatial-channel decoupled downsampling, and rank-guided block design, which greatly reduce the computational redundancy and achieve high efficiency. We further introduce the large-kernel convolution and innovative partial self-attention module, which effectively enhance the performance under low cost. 3. Based on the above approaches, we introduce YOLOv10, a new real-time end-to-end object detector. Extensive experiments demonstrate that our YOLOv10 achieves the state-of-the-art performance and efficiency trade-offs compared with other advanced detectors.

Limitation. Due to the limited computational resources, we do not investigate the pretraining of YOLOv10 on large-scale datasets, e.g., Objects365 [52]. Besides, although we can achieve competitive end-to-end performance using the one-to-one head under NMS-free training, there still exists a performance gap compared with the original one-to-many training using NMS, especially noticeable in small models. For example, in YOLOv10-N and YOLOv10-S, the performance of one-to-many training with NMS outperforms that of NMS-free training by 1.0% AP and 0.5% AP, respectively. We will explore ways to further reduce the gap and achieve higher performance for YOLOv10 in the future work.

Broader impact. The YOLOs can be widely applied in various real-world applications, including medical image analyses and autonomous driving, etc. We hope that our YOLOv10 can assist in these fields and improve the efficiency. However, we acknowledge the potential for malicious use of our models. We will make every effort to prevent this.

Neur IPS Paper Checklist

Question: Do the main claims made in the abstract and introduction accurately reflect the paper s contributions and scope? Answer: [Yes] Justification: We accurately present the contribution and scope of our paper in the abstract and introduction. Guidelines:

The answer NA means that the abstract and introduction do not include the claims made in the paper. The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations

Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We discuss the limitations of our work in the appendix. Guidelines:

The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. The authors are encouraged to create a separate "Limitations" section in their paper. The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory Assumptions and Proofs

Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes]

Justification: We provide the full set of assumptions and the complete proof in the paper and the appendix.

Guidelines:

The answer NA means that the paper does not include theoretical results. All the theorems, formulas, and proofs in the paper should be numbered and crossreferenced. All assumptions should be clearly stated or referenced in the statement of any theorems. The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. Theorems and Lemmas that the proof relies upon should be properly referenced.

4. Experimental Result Reproducibility

Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?

Answer: [Yes]

Justification: We present the detailed information for reproducing the results in the paper and the appendix.

Guidelines:

The answer NA means that the paper does not include experiments. If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. While Neur IPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.

5. Open access to data and code

Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?

Answer: [Yes]

Justification: As shown in the abstract, we provide the code at https://github.com/ THU-MIG/yolov10 and the data is available at https://www.cocodataset.org/. Besides, we provided sufficient information in the paper and appendix to reproduce the experimental results.

Guidelines:

The answer NA means that paper does not include experiments requiring code. Please see the Neur IPS code and data submission guidelines (https://nips.cc/ public/guides/Code Submission Policy) for more details. While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). The instructions should contain the exact command and environment needed to run to reproduce the results. See the Neur IPS code and data submission guidelines (https: //nips.cc/public/guides/Code Submission Policy) for more details. The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.

6. Experimental Setting/Details

Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?

Answer: [Yes]

Justification: We provide all the training and test details in the paper and appendix.

Guidelines:

The answer NA means that the paper does not include experiments. The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. The full details can be provided either with the code, in appendix, or as supplemental material.

7. Experiment Statistical Significance

Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?

Answer: [No]

Justification: We did not report the error bars because the training of YOLO on COCO is stable and consistent.

Guidelines:

The answer NA means that the paper does not include experiments. The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.

The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) The assumptions made should be given (e.g., Normally distributed errors). It should be clear whether the error bar is the standard deviation or the standard error of the mean. It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.

8. Experiments Compute Resources

Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?

Answer: [Yes]

Justification: We provide the information on the computer resources in the implementation details in the appendix.

Guidelines:

The answer NA means that the paper does not include experiments. The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn t make it into the paper).

9. Code Of Ethics

Question: Does the research conducted in the paper conform, in every respect, with the Neur IPS Code of Ethics https://neurips.cc/public/Ethics Guidelines?

Answer: [Yes]

Justification: Our work conforms with the Neur IPS Code of Ethics.

Guidelines:

The answer NA means that the authors have not reviewed the Neur IPS Code of Ethics. If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).

10. Broader Impacts

Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?

Answer: [Yes]

Justification: We discuss the impacts in the appendix.

Guidelines:

The answer NA means that there is no societal impact of the work performed.

If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).

11. Safeguards

Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?

Answer: [NA]

Justification: Our work poses no such risks.

Guidelines:

The answer NA means that the paper poses no such risks. Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.

12. Licenses for existing assets

Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?

Answer: [Yes]

Justification: We used the publicly available dataset in our work. We cited the corresponding papers in the paper.

Guidelines:

The answer NA means that the paper does not use existing assets. The authors should cite the original paper that produced the code package or dataset. The authors should state which version of the asset is used and, if possible, include a URL. The name of the license (e.g., CC-BY 4.0) should be included for each asset. For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.

If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. If this information is not available online, the authors are encouraged to reach out to the asset s creators.

13. New Assets

Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?

Answer: [NA]

Justification: Our work does not release new assets.

Guidelines:

The answer NA means that the paper does not release new assets. Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. The paper should discuss whether and how consent was obtained from people whose asset is used. At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.

14. Crowdsourcing and Research with Human Subjects

Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?

Answer: [NA]

Justification: Our work does not involve crowdsourcing nor research with human subjects.

Guidelines:

The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. According to the Neur IPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.

15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects

Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?

Answer: [NA]

Justification: Our paper does not involve crowdsourcing nor research with human subjects.

Guidelines:

The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.

We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the Neur IPS Code of Ethics and the guidelines for their institution. For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.